perftest
perftest copied to clipboard
configure: Cleanup CUDA configury
The current approach to enabling CUDA support in the bandwidth benchmarks with an environment variable that takes that path to an include file is extremely fragile. As a low-hanging fruit, this commit brings the CUDA configuration closer to ROCm, but relying on a flag to enable the feature and AC_SEARCH_LIBS to update lib flags, or on --with-cuda allowing a user to provide a custom installation path. Ideally, we should rely on dlopen()'ing the symbols we need from these libraries so a single build of perftest can work on systems with and without CUDA SDK, but that's to come later.
Signed-off-by: Raghu Raja [email protected]
It is hard to tell who is maintaining perftests these days, but tagging @HassanKhadour who made the last stable release for review. I do not have triage permissions on this repo to tag reviewers.
By removing the include of the 'CUDA_PATH' in perftest_parameters.h (line 70) the commit is breaking the backward compatibility.
Indeed, I will fix that up. I am in favor of completely removing it if we can.. given the explicit configure warning (which we can turn into an error) and perhaps a call out in the changelog when this makes it into a release.
Does perftest.spec need a change as well?